Trust “done” again.

Yes, and it is now two programs rather than one. The coding agent runs in your terminal with no account at all, and the gate verifies your code locally, forever — the full verification loop, signed receipts included. Neither checks a licence. What you supply is the model: connect a Claude, ChatGPT or Copilot subscription you already pay for, or your own API key.

No. Relay never resells model usage and never marks it up. You connect your own Anthropic, ChatGPT or GitHub Copilot subscription, or an API key from any of about thirty providers, and that account is billed by that provider exactly as it is today. The agent will not run a turn until one is connected — that is the one thing you have to bring.

Start the agent under relayevals supervise. Relay runs the test command itself and signs the result with a key the agent never holds, so the receipt records that Relay produced it rather than that the agent said so. A repository can then refuse anything else, and can name exactly which machines it will accept receipts from at all. That is local, works offline and is the same on every plan. What Team adds is that the same verdict is enforced at the merge gate and recorded.

The credentialed gate: a CI signing credential via GitHub OIDC, PR check records only your CI can write, the org baseline enforced at the gate — a repository cannot quietly weaken it, and the analytics and audit history those records fill. $12 per developer per month billed yearly ($144 per seat), or $15 month-to-month.

Cancel, and Team runs until the period you paid for ends. A failed card does not cut you off on the first attempt — the plan is carried while your provider retries, and only a revoked subscription, refund or chargeback ends it. Either way, local verification never stops and every receipt you hold stays verifiable — what lapses is the CI signing credential.

Yes — that is the admin role. An admin registers repositories and publishes the organization baseline; the owner keeps billing, the roster and deletion. What does not exist yet is governance of the governors: approvals before a policy change lands, named reviewer groups, SSO. If you need those today, Relay is not the answer yet and we would rather say so.

Developers, not dashboards. Reviewers, viewers and bots are not seats. We don’t meter this — the count you set in Polar is the one that bills, pre-filled with your accepted developers, and the members page shows any drift.

No. Coverage, receipts and the answer are identical on every plan. Import-graph tests fail the build if an entitlement check ever reaches a verdict path.

Two ways, both flat per seat: $144 per seat per year — $12 a month — on one invoice, or $15 per seat month-to-month which you can stop any month. Never usage-based, never credits, and billed through Polar either way.

The verdict runs locally. Shared memory travels through your own git and never crosses Relay’s servers. CI reports carry counts and codes only — no filenames, no source. The agent is the same shape: your code and prompts never reach us. If you sign a machine in, it reports what a session cost — a session id, counts and a number of micro-dollars, and no field that could carry a path, a prompt, a branch, a repository or a task. It is on once you sign in, and /account reporting off stops it.

The gate is the verdict’s exit code, which runs without us. Reporting the decision can fail without failing your build.

Partly, and deliberately. The drive standard and shared lessons travel through your own git, and the CLI never phones home to check a licence — those run the same on every plan and always will. What a plan buys is the recorded gate: the CI signing credential, and the analytics, audit history and org baseline that the records it signs make possible. Team is the deal, not a lock.

Polar is our Merchant of Record and appears as the seller on your statement; refunds and invoices go through Polar.